Most cash forecasts fail for a mundane reason: they are built on closing balances and monthly averages, while payments are decided hour by hour. A multi-currency forecast has to predict not just how much liquidity a business needs, but when and where it will be needed — across currencies, corridors and time zones. Get that wrong and the group either runs short at the worst moment or holds buffers it does not need.

Start with the rhythm of flows

Begin with recurring flows rather than totals. Salaries, supplier runs, card settlement, tax dates and subscription cycles create predictable spikes. Layered on top are agent-driven or API-driven operations, which can arrive continuously and in small amounts. Separating the two helps: scheduled flows can be planned, while continuous flows need a pattern, not a schedule.

Build the model in three layers

  1. A baseline of historical inflows and outflows per currency and corridor, refreshed from the ledger rather than from spreadsheets.
  2. A buffer for timing uncertainty, sized to the corridors with the widest cut-off gaps.
  3. A routing layer that asks, for each payment, whether it needs external liquidity at all — or whether offsetting flows inside the network can settle it first.

The third layer is where forecasting becomes cheaper. If opposite flows cancel before they leave, the forecast no longer has to fund every corridor at peak. Only the net remainder needs real external liquidity.

Use evidence, not estimates

A forecast is only as good as its inputs. Operations recorded in a double-entry ledger and attributed to a principal, an agent and a policy version give treasury an auditable history to model against. Around ten currencies, per-unit reporting and signed, verifiable documents make reconciliation a routine step instead of a monthly project. Where balances sit with licensed partners, the forecast still works — visibility does not require ownership.

In short

  • Forecast the timing and location of need, not just the total.
  • Separate scheduled flows from continuous, agent-driven ones.
  • Netting shrinks the amount the forecast must actually fund.
  • Ledger evidence beats spreadsheet estimates as an input.

See how netting changes what treasury has to forecast → /protocol/